光计算机互连的机遇：元平台视角.pdf

《光计算机互连的机遇：元平台视角.pdf》由会员分享，可在线阅读，更多相关《光计算机互连的机遇：元平台视角.pdf（11页珍藏版）》请在三个皮匠报告上搜索。

1、OCP Global Summit October 18,2023|San Jose,CADrew Alduino,Meta Platforms AI SystemsOPPORTUNITIES FOR OPTICAL COMPUTER INTERCONNECTS META PERSPECTIVEParameterUnitsMinMaxCommentsData rate per laneGbps#lanes/link#Aggregate BW/linkGbps12,00016,0001.5-2TBpsLinear BW densityGbps/mm1,1001,50011mm shoreline

2、Areal BW densityGbps/mm2Energy efficiency single direction linkpJ/bit3Energy efficiency with external light sourcepJ/bitTotal Energy EfficiencypJ/bit4Reachmeters30100Latency(no FEC)nanosec2026Across many applications/services and at very large scales,driving a large portion of our overall infrastruc

3、ture both HW and SW Ranking and recommendation Video,Ranking,Search.Content understanding Computer vision,Speech,Translation,NLP,Video.From data centers to the edge-AI Use Cases at MetaKeypoint SegmentationAugmented Reality with Smart CameraIO Protocol&Component Scaling is dramatically(32x)falling b

4、ehindAI Model Size ScalingAI Model TrendsAI Model Size 2x/4 monthsComponent Capacity2x/24 monthsFrom:Amir Golani etal https:/ Scaling of Peak hardware FLOPS,and Memory/Interconnect BandwidthGPU Communications:IO Use CasesGPU-CPU IOLargely PCIe based(512Gbps 1Tbps)Grace Hopper is changing this(900GBp

5、s/7.2Tbps)GPU-GPU IOScale-up&Scale-out regimes=Back-end fabricMulti-end point,1TBps scaling at 50%gen over gen?Currently supported with tens of 400G/type pluggable modulesGPU-Memory IOHBM communication fast,low latency,“low capacity”memoryPoint to point,multi-end point(?)HBM BW scaling 50%gen over g

6、en?(1/1.5/2.25TBps)BW,BW-density,economics,power,latency are challenging for optical IO and will require innovationIntra-board:GPU-Memory?,Disaggregation?,GPU-CPUCluster Architecture“Requirements”Optical InterconnectToR-SpineSpineSwitchRackSwitchGrand TetonSer